AI tools

The best open-weight language models, according to hackers

39 open-weight language models recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.

• Most recommendedQwen 3.8 27B1173 people recommend it
• Built to lastQwenowners report a median of 0.5 years
• Least regretNemotron 3 Super20 recommend, 3 criticize

All open-weight language models

• Qwen 3.8 27BAn open-weights model that runs on a single high-end GPU and nears frontier coding quality, though it overthinks, burns tokens and struggles with complex tasks.1173recommend
• Z.ai GLMAn open-weight coding model close to frontier quality at a fraction of the cost, though critics cite slow, outage-prone hosting, erratic output and expensive self-hosting hardware.1043recommend
• KimiMoonshot's open-weight models rival top closed models at coding and research for a fraction of the price, though they can be slow and overthink.931recommend
• Gemma 4Google's open-weight model family runs well on local hardware with strong quality and token efficiency for its size, though critics cite unreliable tool calling and weaker coding than Qwen.713recommend
• QwenThe go-to open-weight model family for running capable coding and general AI locally on accessible hardware, though it falls short of frontier models like Claude Opus.474recommend
• Qwen3.6-35B-A3BA fast mixture-of-experts open model that runs well on consumer hardware and is surprisingly capable for its size, though it falls well short of frontier models on complex tasks.373recommend
• gpt-ossOpenAI's open-weight models are fast, cheap or free to run, and capable enough for agents, tool calls and basic tasks; critics call them heavily censored and weak on harder work.196recommend
• GLM 5.3 FlashAn MIT-licensed open-weight model with near-frontier coding ability at a fraction of the cost, runnable locally or via many hosts; critics distrust Z.ai's terms and China ties.192recommend
• MiniMax M2An open-weight coding model that runs locally and comes close to Sonnet and Opus at a fraction of the price; critics cite slowness and heavy token use.154recommend
• Qwen 3.8 Flash NextAn open-weight model that runs fast on consumer GPUs and Macs, with coding ability some call near frontier models; it overthinks by default and degrades at low quantization.147recommend
• Meta LlamaRecommended as the model family that kicked off open-weight AI and still runs cheaply on local hardware, though critics say it has fallen behind and Llama 4 disappointed.139recommend
• Bonsai 27BPeople recommend it as a 27B-class model that fits in little memory and runs on laptops and phones. The common caveat is that it weakens on long tasks and on vision.89recommend
• DeepSeek modelsPeople recommend DeepSeek's open-weight models for strong performance at low cost and self-hosting. Critics mainly point to refusals on political questions.75recommend
• DeepSeek V3Recommended as a cheap, open-weight model that gets close to closed frontier models on many tasks. Common caveat: slow with large contexts.59recommend
• Nvidia NemotronPraised for open training pipelines, data, tooling and free licensing. Critics call results inconsistent and releases lackluster next to Qwen and other Chinese models.50recommend
• Olmo by Ai2Ai2's fully open language model releases weights, training data, code and checkpoints, so people recommend it as the most transparent option, though it trails top closed models.47recommend
• GLM-4.7-FlashRecommended as a fast, capable local model for coding and tool calling, roughly on par with mid-size rivals. Critics find it weaker than bigger models and report chat template bugs.46recommend
• Thinking Machines InklingThinking Machines' open-weights model, praised as a strong American debut and for its fast, pleasant tone, but critics say it trails rivals on benchmarks.42recommend
• IBM GraniteSmall open-weight models from IBM that people run locally, praised for speed, efficiency and strong results for their size. Critics say they trail rival models on benchmarks and lack reasoning training.37recommend
• OrnithLocal models that people find fast and useful for coding, often ahead of earlier Qwen releases. Critics say coding results are uneven and benchmark claims are unproven.37recommend
• Kolibri by Aleph AlphaOpen-weight Apache 2.0 model from Aleph Alpha that people praise for openness, offline-friendly use and strong benchmarks. Critics say it trails Qwen models, is heavy to run locally, and overthinks.33recommend
• Tencent Hy3Recommended as a cheap, capable open-weights model that punches above its size for coding and writing. Caveats: slow responses, rate limits and errors on OpenRouter.25recommend
• LFM2.5-24B-A2B by Liquid AIFast, efficient open-weight model for on-device agents; critics say it lags Qwen on knowledge and coding, and the license is confusing.23recommend
• Nemotron 3 SuperOpen-weight Nvidia model that people run locally and praise for fast inference and open training data. Critics say it trails Qwen models, especially at coding.20recommend
• Kimi K2.7People recommend it for strong coding and bug finding at a low cost compared with other models. Critics say it needs more guidance and uses many tokens.18recommend
• ApertusFree fully open model with open weights, data and training details, valued for copyright safety and research. Critics say it trails US and Chinese models.17recommend
• Arcee Trinity LargePraised as a rare fully open-weights American model with strong value; critics say it is undertrained and weak on agentic tasks.16recommend
• Mistral 7BFree open-weights model people run locally for fast, offline use, especially on JavaScript and multilingual tasks. Critics call it dated next to newer models and unreliable for production.13recommend
• Poolside Laguna XS.2Recommended for fast local use on CPU and for web research; critics say it trails rivals on benchmarks and is broken in places.12recommend
• QwopusPeople recommend Qwopus, a Qwen fine-tune in Opus style, for solid local coding and grunt-work tasks across multiple files. Critics say its marketing overstates it and that results vary between runs.10recommend
• GLM-4.5 AirFree open-weight model that people run locally and rate highly for coding and general use. Some report weaker performance on certain tasks.9recommend
• Nvidia Nemotron 3.5 LightningFast local model praised for decent code and an open training pipeline; critics say it overthinks and is weaker than comparable 27B and 35B models.9recommend
• Sarvam LLMIndian open-weight language model praised for being trained from scratch, cheap, and strong on local languages. Critics say it hallucinates and lacks transparency about its claims.8recommend
• Nvidia Nemotron UltraFree open-weight Nvidia model that people find fast and competitive with top models, though one critic found it often down on OpenRouter.8recommend
• Qwen-AgentWorld 35B A3BQwen's post-trained 35B A3B agent model that people find stronger than Qwen 3.6 for agentic work and developer help. Critics note reasoning loops and questionable world-model branding.8recommend
• K2 HorizonOpen model family praised for sharing training data and recipes and for its 7B model; critics say headline benchmark claims do not match self-reported results and larger models underperform.8recommend
• Falcon models by TIIFree open-weights language models from TII that people call credible and competitive, though one critic says they are not competitive with leading models.6recommend
• Ernie by BaiduBaidu's open weight model family, recommended as one of the good Chinese AI models; critics note answers stay locked on political topics.6recommend
• Thinking Machines 975B open-weights LLMLarge open-weights model from a new lab, with vision input; people credit the open release, though it trails some open models on benchmarks.6recommend